|
|
| import os
|
| import sys
|
| import argparse
|
| import json
|
| import datetime
|
| import cv2
|
| import numpy as np
|
| import torch
|
| import lpips
|
| from torchvision import transforms
|
| import torch.nn.functional as F
|
| from PIL import Image, UnidentifiedImageError
|
|
|
| def verify_image(path, exts=('.png', '.jpg', '.jpeg', '.webp')):
|
| """检查文件存在、非空、扩展名合法，并能被 PIL 打开。"""
|
| if not os.path.isfile(path):
|
| return False, f'文件不存在：{path}'
|
| if os.path.getsize(path) == 0:
|
| return False, f'文件为空：{path}'
|
| if not path.lower().endswith(exts):
|
| return False, f'不支持的格式：{path}'
|
| try:
|
| img = Image.open(path)
|
| img.verify()
|
| except (UnidentifiedImageError, Exception) as e:
|
| return False, f'无法读取图像：{path} （{e}）'
|
| return True, ''
|
|
|
| def load_tensor(path):
|
| """按原脚本方式载入并归一化到 [-1,1] 的 Tensor"""
|
| img = cv2.imread(path, cv2.IMREAD_COLOR)
|
| if img is None:
|
| raise RuntimeError(f'cv2 读取失败：{path}')
|
| img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
| t = transforms.ToTensor()(img) * 2 - 1
|
| return t.unsqueeze(0)
|
|
|
| def histogram_intersection(a, b, bins=256):
|
| """计算两张图 RGB 通道直方图的平均交集比率"""
|
| inters = []
|
| for ch in range(3):
|
| h1 = cv2.calcHist([a], [ch], None, [bins], [0,256]).ravel()
|
| h2 = cv2.calcHist([b], [ch], None, [bins], [0,256]).ravel()
|
| if h1.sum() > 0:
|
| h1 = h1 / h1.sum()
|
| if h2.sum() > 0:
|
| h2 = h2 / h2.sum()
|
| inters.append(np.minimum(h1, h2).sum())
|
| return float(np.mean(inters))
|
|
|
| if __name__ == "__main__":
|
| p = argparse.ArgumentParser(description='自动化风格迁移效果检测脚本')
|
| p.add_argument('--groundtruth', required=True, help='真实值目录路径')
|
| p.add_argument('--output', required=True, help='风格化后图像路径')
|
| p.add_argument('--lpips-thresh', type=float, default=0.5, help='LPIPS 阈值 (>= 通过)')
|
| p.add_argument('--hi-thresh', type=float, default=0.6, help='HI(直方图交集) 阈值 (>= 通过)')
|
| p.add_argument('--result', required=True, help='结果 JSONL 文件路径，追加模式')
|
| args = p.parse_args()
|
|
|
|
|
| content_path = os.path.join(args.groundtruth, 'images.jpg')
|
| style_path = os.path.join(args.groundtruth, 'style.jpg')
|
|
|
| process = True
|
| comments = []
|
|
|
|
|
| for tag, path in [('content', content_path), ('style', style_path), ('output', args.output)]:
|
| ok, msg = verify_image(path)
|
| if not ok:
|
| process = False
|
| comments.append(f'[{tag}] {msg}')
|
|
|
|
|
| lpips_pass = hi_pass = False
|
| lpips_val = hi_val = None
|
| if process:
|
| try:
|
|
|
| img_c = load_tensor(content_path)
|
| img_o = load_tensor(args.output)
|
|
|
| _, _, h0, w0 = img_c.shape
|
| _, _, h1, w1 = img_o.shape
|
| nh, nw = min(h0,h1), min(w0,w1)
|
| if (h0,w0)!=(nh,nw):
|
| img_c = F.interpolate(img_c, size=(nh,nw), mode='bilinear', align_corners=False)
|
| if (h1,w1)!=(nh,nw):
|
| img_o = F.interpolate(img_o, size=(nh,nw), mode='bilinear', align_corners=False)
|
|
|
| loss_fn = lpips.LPIPS(net='vgg').to(torch.device('cpu'))
|
| with torch.no_grad():
|
| lpips_val = float(loss_fn(img_c, img_o).item())
|
| lpips_pass = lpips_val >= args.lpips_thresh
|
|
|
|
|
| img_s = cv2.imread(style_path, cv2.IMREAD_COLOR)
|
| img_o_cv = cv2.imread(args.output, cv2.IMREAD_COLOR)
|
| hi_val = histogram_intersection(img_s, img_o_cv)
|
| hi_pass = hi_val >= args.hi_thresh
|
|
|
| comments.append(f'LPIPS={lpips_val:.4f} (>= {args.lpips_thresh} → {"OK" if lpips_pass else "FAIL"})')
|
| comments.append(f'HI={hi_val:.4f} (>= {args.hi_thresh} → {"OK" if hi_pass else "FAIL"})')
|
|
|
| except Exception as e:
|
| process = False
|
| comments.append(f'指标计算出错：{e}')
|
|
|
|
|
| result_flag = (process and lpips_pass and hi_pass)
|
| entry = {
|
| "Process": process,
|
| "Result": result_flag,
|
| "TimePoint": datetime.datetime.now().isoformat(sep='T', timespec='seconds'),
|
| "comments": "; ".join(comments)
|
| }
|
| print(entry["comments"])
|
| os.makedirs(os.path.dirname(args.result) or '.', exist_ok=True)
|
| with open(args.result, 'a', encoding='utf-8') as f:
|
| f.write(json.dumps(entry, ensure_ascii=False, default=str) + "\n")
|
|
|
